Artificial intelligence and customers’ intention to use

Research background: Robo-advisory is a modern and rapidly developing area of implementing artificial intelligence to support customer decision-making. The current significance of robo-advisory to the financial sector is minor or marginal, and boils down to formulating recommendations and implementing investment strategies. However, the ongoing digital transformation of the economy leads us to believe that in the near future this technology will also be much more widely used with banking products. This makes it necessary for banks and other financial institutions to be prepared to offer this


Introduction
In recent decades, there has been a noticeable increase in the importance of the financial sector for both business entities and individual consumers.Financial institutions have skillfully expanded and tailored their services to meet the needs of a strongly growing and globalizing economy (Śliwiński, 2023;Purewal & Haini, 2022).They have also recognized the attractiveness of the retail customer segment by offering financial services based on modern information and communication technologies (Gallego-Losada et al., 2023;Mehdiabadi et al., 2022).
The use of electronic channels of communication with customers reduced the constraints of time and place of service delivery, and reduced the costs and barriers related to accessing financial services (Hussain et al., 2023).Consumers gained the ability to self-deliver standardized financial services, but the use of advice still required contact with bank or brokerage staff.The next phase of digital transformation, manifesting itself in the wider use of artificial intelligence in financial services, enabled progress also in the area of advice (Ahmed et al., 2022;Rodrigues et al., 2022;Doumpos et al., 2023).Drawing on developments in FinTech, automated advice is finding its way into decision-making for an increasingly broad spectrum of financial instruments, representing a promising as well as a socially responsive area of business practice for the digital age (Harasim, 2021;Małkowska et al., 2021;Adamek & Solarz, 2023;Waliszewski et al., 2023).
Robo-advisory is defined as an automated digital investment advisory programme (U.S. Securities and Exchange Commission, 2017).The term robo-advisors is also understood as online platforms that use artificial intel-ligence to process information on clients' investment preferences (Maume, 2021).In robo-advisory, artificial intelligence fully or partially replaces interaction with staff (D'Acunto et al., 2019).In the case of fully automated advice, an advanced degree of artificial intelligence algorithms enables it to make decisions on the creation and execution of investment strategies.In other cases, the software is responsible for preparing investment recommendations, as well as selecting assets for the portfolio, but investment decisions are still made by clients after consultation with a personal financial advisor (Helms et al., 2021;Maume, 2021).To make accurate recommendations, financial intermediaries offering robo-advisory services aim to identify clients' levels of financial literacy, their financial needs and goals, and their attitude towards investment risk.In doing so, they use the information contained in online declarations (Alsabah et al., 2021), as well as those obtained by analysing data extracted from customers' bank and investment accounts (Better Finance, 2020;Day et al., 2018).
Nowadays, regardless of the country analysed, robo-advisory has little or marginal relevance in the financial advice market.This is evidenced by the relatively small value of assets under management and the equally minute number of users.Statista (2023) data shows that in 2022 assets under management in the robo-advisors segment reached an amount of 2.4 trillion U.S. dollars globally.Projections for 2027 show a significant increase in value to 4.5 trillion U.S. dollars, which should be linked to the changes that societies and economies are experiencing in the era of digital transformation.The expected dynamic development of robo-advisory is the result of its numerous advantages, which include, among others: objectivity of decisions, resistance to emotions, 24/7 availability (D'Acunto & Rossi, 2021, pp. 725-749).Moreover, this solution may become very common because it is indicated as suitable for people with little experience in investing and low incomes (Waliszewski, 2022;Piotrowski, 2022;Warchlewska & Waliszewski, 2020).
As the popularity of robo-advisory grows, the number of publications devoted to this topic increases.However, most works are dominated by a narrow understanding of this phenomenon, which is usually reduced to investing.Meanwhile, the expected increase in the use of robo-advisory in financial services, going beyond the currently dominant area of financial investment (European Supervisory Authorities, 2015), indicates the need to introduce a broader definition of robo-advisory.For this reason, our paper defines robo-advisory as automated advice on investing, saving, and ob-taining finance using artificial intelligence technologies to make recommendations or relevant decisions, based on an analysis of client and economic data.
In the literature on the subject, there are few works devoted to the analysis of factors shaping the attitude towards robo-advisory.However, most of them focus on a relatively narrow group of determinants in relation to the technological and operational aspects of the service (Sabir et al., 2023;Nguyen et al., 2023) and the issue of trust in this technology (Yi et al., 2023;Nourallah et al., 2023) with less attention paid to consumer characteristics and experiences.
The aim of this study is to identify factors that significantly influence consumers' intentions to use robo-advisory in bank services.The research carried out will help answer the research question as to which factors, related to consumers' technological sophistication or experience in using financial services, are decisive for the acceptance of robo-advisory.Gaining knowledge about consumers' experiences, expectations and preferences may serve to increase the effectiveness of banks' promotional activities and broaden the acceptance of robo-advisors' services.
The work presented here makes a significant contribution to roboadvisory acceptance research in a number of areas.The research results provide knowledge about the importance of individual factors in the acceptance of robo-advisory.Compared to other works, our study expands the scope of analyzed variables.We go beyond the technological and operational aspects of the service and the issues of trust in this technology and also pay attention to factors related to customers' relationships with banks, including their experience in using electronic banking services and services using artificial intelligence technologies, as well as financial consulting.The paper is one of the few to present research on the use of robo-advisory in financial services, going beyond the realm of financial investments.Such a solution is a response to changes observed in the market regarding the practical application of robo-advisory by banks and other financial market institutions.It should be noted that our study is conducted based on original data collected for this study.Due to the fact that Poland is characterized by a high level of development of the financial sector, comparable to many developed countries, the conclusions received can, to a large extent, be treated as universal and valid also for other modern countries.The originality of the work is also reflected in the research methods used.To the authors' knowledge, this is the first study on robo-advisory acceptance to use a multilevel logit model and modern machine-learning predictive algorithms.
In the later part of the paper, a literature study is conducted on roboadvisory acceptance factors.This is followed by a depiction of how the primary material was obtained, a statistical description of the variables used in the analysis, and the research methods employed.The next section demonstrates and discusses the results obtained using multilevel logit model estimation and methods based on machine learning.The paper is closed by conclusions and recommendations for application.

Literature review
The literature studies conducted by authors reveal a small number of works addressing the acceptance of robo-advisory among consumers.The analyses focused on various groups of factors, which will be described below.
The most frequently studied determinants related to the technological and operational aspects of robo-advisory.Nourallah (2023), Milani (2019), Rühr et al. (2019), Rühr (2020), Gan et al. (2021) and Gerlach and Lutz (2021) used the Unified Theory of Acceptance and Use of Technology (Venkatesh et al., 2003) in order to examine the relevance of key constructs such as: performance expectancy, effort expectancy, social influence, and facilitating conditions.Rogers' Innovation Diffusion Theory (2003) was used in the work of Fan and Chatterjee (2020).This theory focuses on product characteristics and considers factors such as: relative advantage, complexity, divisibility, compatibility, and observability.Davis' Technology Acceptance Model (1989), using determinants including perceived ease of use and perceived usefulness, provided the theoretical foundation for the research conducted in the works of Belanche et al. (2019) and Seiler and Fanenbruck (2021).The Service Robot Acceptance Model (sRAM) originally developed by Wirtz et al. (2018) is also applicable to research on robo-advisory acceptance in bank services.This model develops the Technology Acceptance Model by adding social and relational variables as drivers of service robots' acceptance.The considerations carried out in the work of Fernandes and Oliveira (2021) refer to this model.
Several works focused on the issue of trust in companies providing robo-advisory services, as well as in the technology used in this service itself.
The work of Nourallah (2023) highlights the importance of cultural differences for initial trust in robo-advisory.Lourenço et al. (2020) showed a positive impact of consumers' perceptions of trust and expertise of the firm providing the automated advice and the satisfaction with the consumeronline tool interaction on the acceptance of robo-advisory.They further proved that expertise influences the acceptance of automated advice more than trust.Yi et al. (2023) pointed out the importance of trust, but also financial knowledge and usability perception.Bruckes et al. (2019) found that a high intention of using robo-advisors was positively influenced by structural assurance and trust in analyzed technological solution.Moreover, it was shown that trust in banks predicted trust in robo-advisors.Ngo-Ye et al. (2018) indicated that the level of trust in robo-advisors is determined by factors such as security, information quality, and interface design.Cheng et al. (2019) found that reputation, information quality, and service quality have a positive correlation with trust in the provider, while factors such as government regulation, attitudes towards artificial intelligence, and supervisory control positively influence trust in robo-advisor technology.Gan et al. (2021) showed that performance expectancy, social influence, and consumer's trust towards robo-advisors are positive and significantly related to the behavioral intention to adopt financial robo-advisors.In turn, the results of the study by Nourallah et al. (2023) indicate that trust in financial robo-advisors depends on the level of knowledge about the technology used in robo-advisory, especially obtained from social media, and on the decision-making style -intuitive or rational.
Some works identify the demographic and socio-economic factors as important determinants of the acceptance of robo-advisory.Milani (2019) points to the level of education, investing experience, and understanding of robo-advisors.Fan and Chatterjee's (2020) study yielded that the need to free up time, higher risk tolerance, higher subjective financial knowledge, and higher amounts of investable assets are positively related to the acceptance of robo-advisory by individual investors.The work of Liao et al. (2022) indicates the importance of the level of financial literacy of respondents.On the other hand, a study by Niszczota and Kaszás (2020) indicates that consumer reluctance to adopt robo-advisory is a manifestation of the wider phenomenon of aversion to machines making decisions, particularly for tasks requiring subjective (moral) judgements.The findings of Zhang et al. (2021), suggest that the reason for the preference for human financial advisors may be connected with their higher expertise in comparison to robo-advisors.When considering the reasons for the difficulties of automated advisory proliferation, it is also important to mention the lack of consumer knowledge about the functioning of artificial intelligence algorithms resulting in an increase in information asymmetry.
The last group of works focused on factors based on the comparison of advice using artificial intelligence algorithms to traditional advice provided by employees of banks and investment companies.Phoon and Koh (2018) emphasize the lower costs of asset management and the greater accessibility of financial services, with accessibility being understood both as the possibility to use automated advisors 24/7 and lower minimum investment amounts required to benefit from professional asset management.Kordela (2018) emphasizes more time-efficiency, while Uhl and Rohner (2018) underline the greater efficiency and transparency of a passive investment strategy typically implemented by robo-advisors.In turn, the study by Tao et al. (2021) demonstrates the domination of robo-advisory in risk-adjusted performance.Brenner and Meyll (2020) indicate that, for some investors, the advantage of robo-advisory may be the avoidance of conflicts of interest and investment fraud occurring with human financial advice.In contrast, Glaser et al. (2019, pp. 133-138) andFoerster et al. (2017) highlight less behavioral bias in the consumer-advisor relationship from robo-advisors compared to human advisors.Jung et al. (2019, pp. 405-427) cite another advantage in the form of less emotional decision making but, on the other hand, they also mention disadvantages of robo-advisory such as the lack of personal contact and the non-acceptance of this form of service delivery.
In summary, research into the acceptance of robo-advisory focuses on the technological and operational aspects of the solution.Only limited reference has been made to the demographic and socio-economic determinants of acceptance.The research completely ignores the issue of consumers' experience in using financial services and the determinants relating to the ethical aspects of customer service by banks.Moreover, robo-advisory is generally analyzed in relation to investments in the financial market.Among the analysis methods used, the use of a multilevel logit model and machine-learning predictive algorithms was not observed.Going beyond these limitations, the present work extends and complements the research presented above.

Research methods
The paper uses the results of a survey carried out as part of a project on the application of artificial intelligence technology in the banking sector in Poland.This survey took into account the consumers' point of view.The collection of primary data was commissioned to a professional research agency -Interactive Research Center Sp. z o.o.The agency's compliance with the regulations contained in the International ICC/ESOMAR Code ensures high quality of research and its compliance with the principles of ethics.The survey was conducted by means of the computer-assisted telephone interview method using a questionnaire prepared by the authors.The questionnaire included sections such as: demographics, technological sophistication of the respondent, trust towards banks, experience of using banking products, perceptions of artificial intelligence and robo-advisory in banking, and personal data management.The pilot survey carried out in September 2020 allowed the comprehensibility of the questions in the questionnaire to be verified and also provided a basis for modifying some of the questions.The full-scale survey conducted in October 2020 covered a sample of 911 Polish citizens aged 18-65.The selection of respondents for the sample was random-quota.The sample was representative of Polish society in terms of gender, age, and place of residence.
Respondents who were members of the agency's research panel took part in the survey.Participation in the panel, as well as participation in the study, was completely voluntary.Moreover, each respondent had the option to stop the study at any time.The study itself and the data obtained in it were anonymous.Considering the fact that the survey was noninterventional and, moreover, it was not a clinical trial, the Research Ethics Committee of the Faculty of Economic Sciences and Management of the Nicolaus Copernicus University in Toruń decided that ethical consent was not required in the analyzed case.
Table 1 presents the variables used in the analysis and the structure of the responses given by the respondents.The survey questionnaire is available in Appendix 1.The dependent variable Robo Intention refers to respondents' attitudes towards using robo-advisory that supports banking services in the following five years.The explanatory variables relate to the socio-demographic characteristics of the respondents, their attitudes towards modern digital technologies, and their experience in using banking services.
To assess the relationship between explanatory variables, their independence was tested.Due to the fact that some variables under study are nominal and others are ordinal, we applied the Chi-Square test of independence, comparing the frequency of each category for one variable across the categories of the second variable.The results of this study are presented in Tables 2a and 2b.As observed, many analyzed variables are dependent, which may hinder the analysis and adversely affect the formulation of conclusions.
In order to select variables with the highest influence on robo-advisory acceptance in banking services, we applied the econometric multilevel ordered logit model and several tree-based machine learning algorithms.Logit and probit models are among the most popular models for binary response.In our case, the dependent variable takes five ordered values hence the multilevel ordered logit model needs to be applied.Multilevel ordered logit models constitute a simple variation on ordinary logit/probit models, and are usually applied when the dependent variable is a discrete and ordered measurement (Matuszewska-Janica & Witkowska, 2021;Cottrell & Lucchetti, 2022).It is assumed that the ordered response variable Y can take on any of the J + 1 values 0, 1, … , J, and it is supposed that underlying the observed response is a latent variable (1) The "cut points" α < α < ⋯ < α are defined, such that The unknown parameters α have been estimated jointly with the βs via maximum likelihood, using Gretl 2021a (cf.Cottrell & Lucchetti, 2022).
It is crucial to underscore that the selection of a subset of variables from a larger pool of potential variables using any econometric model is generally a complex and contentious issue.Various strategies can be employed for variable selection, yet there are no definitive guidelines on which ones yield the optimal results.For instance, Olejnik et al. (2000) conducted a study comparing the effectiveness of three alternative strategies: the stepwise selection method and two others based on Mallow's !" and Wherry's adjusted R² statistics to determine the final best model.They found that although, overall, the outcomes of all compared strategies were unsatisfactory, the stepwise selection method yielded relatively better results.
In order to indicate significant variables in ordered response models, we applied the backward elimination procedure.It means that we started with a full model with all considered variables and then removed the least significant variables one at a time.We continued this iterative procedure until we obtained variables for which z-statistic was significant at the 5 percent level.To evaluate the final model, an additional assessment was conducted based on its economic interpretation, specifically by scrutinizing the accuracy of the signs of the estimated coefficients for individual variables.This stage serves to validate the research procedure applied and enhances the reliability of the derived conclusions.However, given the aforementioned uncertainties, and to further validate the variables derived from the logit model, this paper opted to complement the study with additional alternative approaches based on machine learning.
In recent years, machine learning algorithms have gained popularity for predictive purposes.They are data-driven, self-adaptive methods requiring very few assumptions about investigated data, which are capable of approximating non-linear relations based on noisy and non-stationary data (Orzeszko, 2021;Fiszeder & Orzeszko, 2021).Among the most popular machine learning algorithms are decision trees and other tree-based methods.Tree-based methods are supervised learning tools used to learn a func-tion that combines a set of variables intended to predict another variable.They partition the feature space into a set of rectangles, and then fit a simple model (like a constant) in each one (Bejger & Fiszeder, 2021;Hastie et al., 2009).Since the response variable is categorical we created classification trees using the standard CART algorithm (Breiman et al., 2017).
Generally speaking, there are two steps of building a classification tree (cf.Hastie et al., 2009;James et al., 2021;Papík & Papíková, 2021): 1. Divide the predictor space -that is, the set of possible values for the majority class in this region.To this end calculate the proportion of class k observations in R ' : where N ' is the number of observations in R ' and I is the indicator function.Finally, the predicted class for this region is We build classification trees in Matlab R2020, where optimal regions R , R , … , R & (corresponding to the nodes of the tree) are indicated in the following steps: 1. Start with all input data, and examine all possible binary splits on every predictor.2. Select a split with the least value of the Gini's diversity index.3. Impose the split.4. Repeat recursively for the two child nodes.
Splitting continues until one of the following stopping rule is triggered: − The node contains only observations of one class, − There are fewer than MinParentSize=10 observations in this node.− Any split imposed on this node produces children with fewer than MinLeafSize=1 observations.− The algorithm splits MaxNumSplits=N-1 nodes (where N is the training sample size).
Apart from a single classification tree we applied several tree-based ensemble models.The idea of ensemble models is to combine two or more models (so-called weak learners) to enable a more robust classification.Applying ensemble models reduces prediction variance and prevents bias from increasing.Moreover, it may be easier to train several simple classifiers and combine them into a more complex classifier than to learn a single complex classifier (cf.Breiman, 2001;Ferreira & Figueiredo, 2012, pp. 35-85).
In our study, we applied four techniques of creating tree-based ensemble models: bagging, adaptive boosting, random undersampling boosting, and totally corrective boosting.Generally, all these procedures improve the predictive performance of weak learners, however the bagging procedure turns out to be a variance reduction scheme and boosting methods primarily reduce the model bias (Bühlmann, 2012(Bühlmann, , pp. 985-1022)).
Bagging (short for bootstrap aggregation) introduced by Breiman ( 1996) is a general approach that uses bootstrapping in conjunction with any regression or classification model to construct an ensemble.Each model in the ensemble is then used to generate a prediction for a new bootstrap sample and these predictions are averaged to give the bagged model's prediction (Kuhn & Johnson, 2013).It has been shown by Bühlmann and Yu (2002) that bagging is a smoothing operation which turns out to be advantageous when aiming to improve the predictive performance of regression or classification trees.In the case of decision trees, the theory in Bühlmann and Yu (2002) confirms that bagging is a variance reduction technique, reducing also the mean squared error (Bühlmann, 2012(Bühlmann, , pp. 985-1022)).
We also applied three algorithms of boosting for multiclass classification: adaptive boosting (AdaBoost.M2, Freund & Schapire, 1997), random undersampling boosting (RUS-Boost, Seiffert et al., 2010) and totally corrective boosting (TotalBoost, Warmuth et al., 2006).Generally, in boosting techniques, a sequential aggregate of base classifier is constructed on weighted versions of the training data, focusing on misclassified samples at each stage of generating classifiers based on the sample weights that are changed according to the performance of the classifier (Tanha et al., 2020;Ferreira & Figueiredo, 2012, pp. 35-85).It means that unlike bagging, which is a parallel ensemble method, boosting methods are sequential ensemble algorithms.Boosting can be viewed as a nonparametric optimization algorithm in function space and has been empirically demonstrated to be very accurate in terms of classification (Bühlmann, 2012(Bühlmann, , pp. 985-1022)).
In order to compute estimates of predictor importance in the constructed tree-based models, we summed changes in the risk due to splits on every predictor and dividing the sum by the number of branch nodes where I is the risk of a parent node, I and I are the risks for child nodes and J KLMNOP is the total number of branch nodes.A node risk is defined as a node impurity weighted by the node probability: where p 4 is the node probability of node i, and E 4 is the node impurity, calculated using the Gini's index, of node R.

Results and discussion
First, we present the results of econometric modeling using multilevel ordered logit model.In Table 3, we summarized the estimation results of the final model obtained as the result of the backward elimination procedure.
All the explanatory variables of this model are significant at the 5 percent level and represent the most important determinants of robo-advisory acceptance.
One can see that all the estimated coefficients have correct signs, i.e. are negative for Unwanted Products and positive for all the other variables.The percentage of cases 'correctly predicted': 38.6%, is not very high, but seems satisfactory, since the number of classes of the response variable Y is 5.
In Figures 1-5 and in Table 4, we present the results of predictor importance estimates (×10 -4 ) for the tree-based algorithms.For each algorithm we created a rank order list of the predictor importance estimates and presented a place of each variable in this ranking (in brackets).As we can conclude, there are slight differences between results from applied algorithms hence -for better interpretability -we additionally calculated the average importance of each variable (Figure 6 and the last column of Table 4).Additionally, as a synthesis of the conducted research, we have showcased a ranking in Table 5, wherein the analyzed predictors were arranged based on the average importance derived from the tree-based algorithms employed.To facilitate a comprehensive comparison across all utilized methodologies, variables identified as significant in the logit model have been highlighted.The findings from research utilizing tree-based algorithms closely align with the conclusions drawn from the application of the logit model.Notably, five variables identified by the logit model ranked within the top six, with an additional two variables also achieving relatively high positions in the presented ranking.The sole exception is the Investment Advisory variable, which, despite its significance in the logit model, did not demonstrate comparable importance from the perspective of the treebased algorithms.
Estimation results of the multilevel ordered logit model for the variable Robo Intention indicate the relevance of three groups of factors.The first group relates to consumers' attitudes towards digital technologies, with a particular focus on artificial intelligence, the second to the ethical aspects of banks' activities, and the third to the experience of receiving advice in the area of financial investments.The averaged results of the analyses made with machine learning algorithms allowed us to prioritize roboadvisory acceptance factors for bank services and group them under the following categories: consumer attitudes towards artificial intelligence, ethical aspects of bank customer services, experience in using modern information and communication technologies, management of consumers' personal data, demographic characteristics, and consumer experience in the financial services area.
Juxtaposing the results obtained by the two groups of research methods helps see the similarities and differences.The most relevant factors for the acceptance of robo-advisory in bank services are those relating to artificial intelligence technology and the business ethics of banks.The high level of significance or validity (depending on the method applied) of the AI Preferences and AI Quality variables indicates that consumers' belief in the benefits of banks' use of artificial intelligence in the form of better knowledge of customer preferences and improved service quality can significantly influence the consumers' decision to use robo-advisory in bank services.The analysis using machine learning algorithms further indicated the importance of a factor relating to the practical application of artificial intelligence in banks, namely Bank AI Experience.Furthermore, in the two groups of research methods, the variable Test New Technology was found to be significant, important for the acceptance of robo-advisory.The predi-lection for testing new technologies is more than simply using digital solutions.It requires the consumer to be open to innovation, to be willing to make the effort and take the risks involved in using a technological innovation such as artificial intelligence, among other things.It is also worth noting that the averaged results of the different machine learning methods indicate a relatively low importance of the Non-banking AI Experience variable.
Another similarity in the results obtained relates to the importance of variables relating to banks' business ethics.These variables were found to be significant in the multilevel ordered logit model, but to a lesser extent than variables relating to the application of artificial intelligence in banking, and slightly less important in machine learning-based methods.However, a deeper analysis of the results obtained using the different research methods reveals significant differences between them.As a result of the estimation of the multilevel ordered logit model, three variables relating to bank ethics proved to be statistically significant.These were Ethics, Unwanted Products, and Data Ethics.In turn, in machine learning methods, almost all factors relating to banks' business ethics were assigned a high level of importance.Furthermore, due to the large number of variables studied, two subgroups were identified.The first concerns the ethical aspects of bank customer service and includes variables such as Lack Complete Information, Unwanted Products, Honest Advisory, Manipulate Information, and Ethics.The second group concerns the ethical behavior of banks in managing personal data.This group is ranked as the fourth most important and includes variables such as Personal Data Use, Data Ethics, Trust, and Data Sharing.
The use of machine learning methods identified a further two groups of relevant variables: experience of using modern ICT and demographic characteristics.The first includes variables such as Internet Use, Mobile Banking Use, Internet Banking Use, Social Media Use, and E-banking Use.This group of variables was ranked as the third most important in terms of its impact on the acceptance of robo-advisory in bank services.Demographic characteristics included variables such as Gender, Age Group, Residence, and Education.This group was ranked as the fifth most influencing on the acceptance of robo-advisory.
Regardless of the machine learning method used, the group of variables identified as experience in the financial services area was ranked as the least important in terms of the intention to use robo-advisory in bank ser-vices.The averaged results of the survey in this group of methods rank the variables Investment Advisory, Loan Advisory, Financial Advisory, Own Investments and Bank Account at the bottom of the list at positions 24-28.Unlike machine learning methods, the multilevel ordered logit model does not allow the indicated variables to be ranked.Estimation results of the multilevel ordered logit model show a lack of statistical significance for the variables Loan Advisory, Financial Advisory, and Own Investments.The fact that there is statistical significance in the Investment Advisory variable for the decision to use robo-advisory in bank services is one of the main differences between the analyses of the survey results obtained using the two different groups of methods.
When comparing the results of the research carried out within the framework of this study with the results contained in the available publications, it is important to note the different ways in which robo-advisory is defined.In contrast to previous literature, where the acceptance of roboadvisory has been analyzed exclusively in the context of financial investments, the authors refer to this form of customer service as the wide range of financial products offered by banks.Turning to a discussion of the impact of selected factors on robo-advisory acceptance, it can be noted that consumer attitudes towards artificial intelligence were already considered in the study by Cheng et al. (2019), but this variable was analyzed in relation to trust in technology and trust in robo-advisory.In contrast, the results presented in this paper indicate that consumer attitudes towards artificial intelligence are an important factor in the acceptance of robo-advisory in bank services.Moreover, the use of machine learning methods allowed us to identify this group of variables as the most important, most influential in the decision to use robo-advisory in bank services.
To the authors' knowledge, no published work to date has analyzed the impact of a bank's business ethics on the acceptance of robo-advisory.The authors' inclusion of variables relating to honesty and transparency in the bank's relationship with its serviced clients fills a research gap.The importance of this achievement is underlined by the fact that the group of variables concerning ethical aspects of bank customer services proved to be the second most influential on the acceptance of robo-advisory in bank services.The results of the study carried out by the authors furthermore point to the high importance of personal data management variables.The effect of privacy on usage intentions of robo-advisors has only been analyzed previously in the work of Seiler and Fanenbruck (2021).
Much of the research to date has focused on determining the impact of trust on the acceptance of robo-advisory.This trust referred to the institution offering robo-advisory (Lourenço et al., 2020) or to the robo-advisory service (Gan et al., 2021;Bruckes et al., 2019).In contrast to the results of previous research, the multilevel ordered logit model applied in this paper found the Trust variable relating to consumer trust in banks statistically insignificant.The average score based on all the machine learning methods used ranked Trust at a distant position (21st) in terms of influence on the decision to accept robo-advisory.Only with the Classification Tree method was this variable ranked as important (7th in the list).
In several earlier studies (Gan et al., 2021;Hohenberger et al., 2019;Fan & Chatterjee, 2020) the respective authors found that respondents' high level of financial literacy positively influenced robo-advisory adoption.Elsewhere, the study by Milani (2019) demonstrated the relevance of investing experience.In the survey carried out by the authors, respondents were not directly asked about their level of financial knowledge, but instead indicated the type and frequency of use of selected financial services.In this way, the authors gained an in-depth knowledge of respondents' actual financial market activity.This fact sets the authors' study apart from previously published work.When analyzing the results obtained, it can be seen that from the perspective of the acceptance of robo-advisory in banks' services, the use of various e-banking channels by consumers turns out to be a relatively important factor.The authors demonstrated that the variables relating to investments in the financial market and the use of advice on various financial products proved to be the least important variables among those included in the study, which is undoubtedly a significant finding.
Research by Brenner andMeyll (2020) andFulk et al. (2018) indicates that robo-advisors tend to be used by younger people, whereas Milani (2019) identified high levels of education as a factor significantly influencing the intention to use robo-advisory.In the study carried out by the authors using the multilevel ordered logit model, variables relating to demographic characteristics were found to be insignificant.The averaged results of the study using different machine learning methods indicated Residence (rank 13) and Age Group (rank 15) as the most important variables in the group in question.The variable Education was ranked 18th in the list, indicating a low level of importance for this factor.

Conclusions
The results of the authors' research using the multilevel ordered logit model and methods based on artificial intelligence technologies have led to a number of significant and interesting conclusions.The most important of these is the finding of little, if any, influence of financial product usage variables on the decision to use robo-advisory in bank services.Indeed, it appears that the fact that consumers use financial advice or have experience in the area of financial investments (for machine learning methods) are not important factors in the acceptance of robo-advisory.This finding leads to the further conclusion that having experience in the area of financial services, as characterizing older people, is not necessarily a decisive factor in the acceptance of robo-advisory in bank services.On the other hand, the lack of or little experience in using financial services, observed more often among young people, does not necessarily constitute a barrier to using robo-advisory.
The relatively high importance of the use of the e-banking channel variables from among the financial services variables indicates that factors relating to ICT are more relevant than those directly connected with the use of financial services.The very high importance of the variables relating to the issue of digital technology use is confirmed by the results obtained using both groups of research methods for the variables determining consumer attitudes towards the use of artificial intelligence technology by banks.It can therefore be assumed that the willingness to accept roboadvisory in bank services is more dependent on curiosity and openness to technological innovation, conviction about the benefits of artificial intelligence, and experience of using AI in bank services than experience in the financial field.
A second noteworthy achievement of the authors is the recognition of the importance of factors relating to the business ethics of banks.The results of the research carried out indicate that consumers' willingness to use robo-advisory in bank services is strongly dependent on the honesty of financial advice and the transparency of the product offered.Also crucial are the proper practices of banks in the area of personal data management and the willingness of consumers to share such data with banks.It can therefore be concluded that consumers' belief that banks follow ethical principles strongly determines the acceptance of robo-advisory.
Consumers are aware that artificial intelligence technologies are largely replacing humans in automated advice.However, due to the practical lack of robo-advisory on offer from Polish banks, it is difficult for them to determine the level of ethics of advice provided using the technologies mentioned.It can therefore be assumed that the expected level of ethical compliance for robo-advisory is determined by the experience of using the available financial services.
The research carried out allowed the authors to formulate conclusions of an applied nature, which are directed at banks.They address the issue of the use of artificial intelligence technology and the ethical aspects of banks' operations.The results of the study should make bank managers realize that the promotion of digital technologies used in online banking and mobile banking is not a strong enough incentive to decide to use roboadvisory.The use of artificial intelligence technology is of decisive importance in this matter, hence the recommendation: Educational activities should be carried out to show the scope of application of artificial intelligence technology in finance and the resulting benefits for consumers.
A second recommendation of a practical nature is the need to pay more attention to the ethical aspects of customer service.Banks, as institutions of public trust, should formulate the terms and conditions of the financial products they offer and deal with their customers in such a way that there is no doubt that compliance with ethical principles is placed above the pursuit of profit.Considering the paramount importance of banks' business ethics for the acceptance of robo-advice, the authors suggest implementing the following recommendation: In the provision of financial services and communication with clients, ethical issues should be taken into account and emphasized to a greater extent.
In an era of digitalization of the customer service process, the proper management of customers' personal data is an extremely important issue.Consumers' use of digital technologies has allowed banks to acquire huge amounts of data, which, due to the digital form, are easy to process.The significant information advantage of banks may raise consumer concerns and limit the acceptance of robo-advice.The third recommendation combines the issue of artificial intelligence, privacy and ethics: It is necessary to identify consumer concerns regarding data processing using digital technologies, with particular emphasis on decisions made by artificial intelligence algorithms in situations requiring ethical judgement.
In the authors' opinion, the fact that the study analyzed data from a single country does not diminish the value of the results obtained, nor does it limit the practical applicability of the conclusions of the study on an international scale.The high quality of the data obtained was achieved thanks to the careful design of the survey and its implementation by a professional research agency.The use of diverse research methods, on the other hand, made it possible to accurately determine the importance of individual factors for the acceptance of robo-advisory.The choice of Poland as a data collection site is justified by the fact that the country has a well-developed and well-functioning financial services sector, and the population actively uses electronic banking channels.As a result, the conclusions of the research conducted in Poland may be universal.
The results of the study, which indicate the importance of artificial intelligence applications in banking and the business ethics of banks, are, according to the authors, the basis for determining the direction of future research.An important subject of research is the use of artificial intelligence algorithms to make decisions in the area of finance, with particular emphasis on decisions containing an ethical component.In addition, it is also worth constructing and examining the effectiveness of predictive models of consumer acceptance of robo-advisory services provided by banks.Doumpos, M., Zopounidis, C., Gounopoulos, D., Platanakis, E., & Zhang, W. (2023). Operational

Figure 1 .Figure 2 .
Figure 1.Predictor importance estimates for the classification tree

Table 4 .
Predictor importance estimates and their place in the ranking (in brackets)

Table 5 .
Ranking of predictors in terms of their average importance The symbols * and ** indicate statistical significance in the multilevel ordered logit model, with pvalues less than 0.05 and 0.01, respectively.